53     int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);  in jsimd_encode_mcu_AC_first_prepare_neon()  local
54     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);  in jsimd_encode_mcu_AC_first_prepare_neon()
55     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);  in jsimd_encode_mcu_AC_first_prepare_neon()
56     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);  in jsimd_encode_mcu_AC_first_prepare_neon()
57     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);  in jsimd_encode_mcu_AC_first_prepare_neon()
58     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);  in jsimd_encode_mcu_AC_first_prepare_neon()
59     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);  in jsimd_encode_mcu_AC_first_prepare_neon()
60     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);  in jsimd_encode_mcu_AC_first_prepare_neon()
71     int16x8_t sign_coefs1 = vshrq_n_s16(coefs1, 15);  in jsimd_encode_mcu_AC_first_prepare_neon()
74     int16x8_t abs_coefs1 = vabsq_s16(coefs1);  in jsimd_encode_mcu_AC_first_prepare_neon()
76     coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));  in jsimd_encode_mcu_AC_first_prepare_neon()
80     int16x8_t diff1 = veorq_s16(coefs1, sign_coefs1);  in jsimd_encode_mcu_AC_first_prepare_neon()
84     vst1q_s16(values_ptr, coefs1);  in jsimd_encode_mcu_AC_first_prepare_neon()
97     int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);  in jsimd_encode_mcu_AC_first_prepare_neon()  local
98     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);  in jsimd_encode_mcu_AC_first_prepare_neon()
99     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);  in jsimd_encode_mcu_AC_first_prepare_neon()
100     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);  in jsimd_encode_mcu_AC_first_prepare_neon()
101     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);  in jsimd_encode_mcu_AC_first_prepare_neon()
102     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);  in jsimd_encode_mcu_AC_first_prepare_neon()
103     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);  in jsimd_encode_mcu_AC_first_prepare_neon()
104     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);  in jsimd_encode_mcu_AC_first_prepare_neon()
133     int16x8_t sign_coefs1 = vshrq_n_s16(coefs1, 15);  in jsimd_encode_mcu_AC_first_prepare_neon()
136     int16x8_t abs_coefs1 = vabsq_s16(coefs1);  in jsimd_encode_mcu_AC_first_prepare_neon()
138     coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));  in jsimd_encode_mcu_AC_first_prepare_neon()
142     int16x8_t diff1 = veorq_s16(coefs1, sign_coefs1);  in jsimd_encode_mcu_AC_first_prepare_neon()
146     vst1q_s16(values_ptr, coefs1);  in jsimd_encode_mcu_AC_first_prepare_neon()
293     int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);  in jsimd_encode_mcu_AC_refine_prepare_neon()  local
294     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);  in jsimd_encode_mcu_AC_refine_prepare_neon()
295     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);  in jsimd_encode_mcu_AC_refine_prepare_neon()
296     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);  in jsimd_encode_mcu_AC_refine_prepare_neon()
297     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);  in jsimd_encode_mcu_AC_refine_prepare_neon()
298     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);  in jsimd_encode_mcu_AC_refine_prepare_neon()
299     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);  in jsimd_encode_mcu_AC_refine_prepare_neon()
300     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);  in jsimd_encode_mcu_AC_refine_prepare_neon()
312       vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs1, 15)));  in jsimd_encode_mcu_AC_refine_prepare_neon()
319     int16x8_t abs_coefs1 = vabsq_s16(coefs1);  in jsimd_encode_mcu_AC_refine_prepare_neon()
321     coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));  in jsimd_encode_mcu_AC_refine_prepare_neon()
323     vst1q_s16(absvalues_ptr, coefs1);  in jsimd_encode_mcu_AC_refine_prepare_neon()
329     uint8x8_t coefs_eq11 = vmovn_u16(vceqq_s16(coefs1, vdupq_n_s16(1)));  in jsimd_encode_mcu_AC_refine_prepare_neon()
344     int16x8_t coefs1 = vld1q_dup_s16(block + jpeg_natural_order_start[0]);  in jsimd_encode_mcu_AC_refine_prepare_neon()  local
345     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[1], coefs1, 1);  in jsimd_encode_mcu_AC_refine_prepare_neon()
346     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[2], coefs1, 2);  in jsimd_encode_mcu_AC_refine_prepare_neon()
347     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[3], coefs1, 3);  in jsimd_encode_mcu_AC_refine_prepare_neon()
348     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[4], coefs1, 4);  in jsimd_encode_mcu_AC_refine_prepare_neon()
349     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[5], coefs1, 5);  in jsimd_encode_mcu_AC_refine_prepare_neon()
350     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[6], coefs1, 6);  in jsimd_encode_mcu_AC_refine_prepare_neon()
351     coefs1 = vld1q_lane_s16(block + jpeg_natural_order_start[7], coefs1, 7);  in jsimd_encode_mcu_AC_refine_prepare_neon()
381       vmovn_u16(vreinterpretq_u16_s16(vshrq_n_s16(coefs1, 15)));  in jsimd_encode_mcu_AC_refine_prepare_neon()
388     int16x8_t abs_coefs1 = vabsq_s16(coefs1);  in jsimd_encode_mcu_AC_refine_prepare_neon()
390     coefs1 = vshlq_s16(abs_coefs1, vdupq_n_s16(-Al));  in jsimd_encode_mcu_AC_refine_prepare_neon()
392     vst1q_s16(absvalues_ptr, coefs1);  in jsimd_encode_mcu_AC_refine_prepare_neon()
398     uint8x8_t coefs_eq11 = vmovn_u16(vceqq_s16(coefs1, vdupq_n_s16(1)));  in jsimd_encode_mcu_AC_refine_prepare_neon()